Someone else's tools, on your canvas
Register a Model Context Protocol server once and every tool it exposes becomes a node you can drop into a workflow. The same node serves two purposes — a step the orchestrator runs, or a capability an agent may choose to use.
Register an MCP server and its tools become canvas nodes — usable as an ordinary step, or granted to an agent.
How a tool reaches your workflow
- The server describes itself — Point at an MCP server and the tool list comes from the server itself, not from a form you fill in. Add a tool on the server and it appears here; nothing has to be re-declared in two places to stay in step.
- A step, or a capability — What makes a tool a tool is the edge that reaches it, not the node. A normal edge runs it in order, like any other step. An edge from an agent grants it instead — the model may call it once, repeatedly, or not at all. No agent is required to use one.
- Parameters as real fields — A tool's parameter schema renders as a labelled form, not one opaque JSON box you have to get right by hand. Each argument can be a fixed value or an expression drawn from earlier steps, so a tool call rarely needs a transformation step in front of it just to rename fields.
- Its own credential, not yours — A provider's bearer token is stored encrypted against that provider alone, so registering a server does not put a credential where an unrelated workflow could pick it up. The tools a provider exposes are the whole of what it can be asked to do.
Platform capabilities
- 2: Ways to use one tool — a step, or an agent's choice
- From the server: Where the tool list and parameters come from
- Encrypted: Provider token, scoped to that provider alone
Do I need an agent to use an MCP tool?
No. Connect a tool with an ordinary edge and it runs as a scheduled step like any other node, with no model anywhere near it. The agent case is the second use, not the only one — which matters if you want a tool's reliability without a model's unpredictability.
What happens when the server adds a tool?
It shows up. The catalogue is read from the server rather than copied into our configuration, so there is no second list to update and no way for the two to disagree about what exists or what arguments it takes.
Can I control which tools an agent may call?
Yes, and that is the only way it gets one. An agent can call exactly the tools you have drawn an edge to — there is no ambient set it inherits from the workspace. Removing the edge removes the capability, and the graph is therefore a readable statement of what the agent may reach.